Throughput and Delay Analysis of the IEEE 802.15.

3
CSMAICA Mechanism Considering the Suspending
Events in Unsaturated Traffic Conditions
Xin Liu
Software School
Fudan University
Shanghai, China
liu_xin@fudan.edu.cn
Abstract-Unlike in IEEE 802.11, the CSMA/CA traffic
conditions in IEEE 802.15.3 are typically unsaturated. This paper
presents an extended analytical model based on Bianchi's model
in IEEE 802.11, by taking into account the device suspending
events, unsaturated traffic conditions, as well as the effects of
error-prone channels. Based on this model we re-derive a closed
form expression of the average service time. The accuracy of the
model is validated through extensive simulations. The analysis is
also instructional for IEEE 802.11 networks under limited load.
Keywords- CSMA/CA; IEEE 802.15.3; MAC; markov chain;
performance analysis
I. INTRODUCTION
The IEEE 802.15.3 [1] is designed for high-rate multimedia
applications in wireless personal area networks (WPAN). It has
defmed a hybrid MAC protocol based on the TDMA and
CSMA mechanisms [2]. The performance analysis of
CSMA/CA is a key issue in IEEE 802.15.3. The corresponding
analysis of IEEE 802.11 [3] can be used as a reference.
However, in IEEE 802.11 networks, all of the traffic is
transmitted by the CSMA/CA mechanism; thus the traffic
condition is saturated in most cases. In contrast, in IEEE
802.15.3 networks most of the traffic consisting of
synchronous data is sent by the TDMA mechanism, and only a
small amount of traffic comprising commands and
asynchronous data is transmitted by the CSMA/CA mechanism;
hence the traffic condition is unsaturated frequently.
The modeling of CSMA/CA in IEEE 802.11 has been a
research focus since the standards were proposed. The work in
[4] gives the theoretical throughput limit of IEEE 802.11 based
on a p-persistent variant. However, it does not consider the
effect of contention window. In [5, 6] Bianchi analyzed the
saturated throughput of IEEE 802.11 DCF based on markov
chains and stochastic processes. Based on Bianchi's model, [7]
re-analyzes the throughput taking into account the frame retry
limit. Some of the subsequent research is diverted on the
analysis of IEEE 802.11e EDCF capturing the contention
window differentiation [8-10]. A key common assumption in
the above works is that each device always has a frame to
transmit. This is unreasonable if the traffic load is limited, such
947
Wenjun Zeng
Dept. of Computer Sciences
University of Missouri
Columbia, USA
zengw@missourLedu
as in IEEE 802.15.3 networks. To our best knowledge, little
analysis is available for unsaturated conditions except in [11-
14]. According to the specifications in [1, 3], whether or not the
buffer has a frame to send after a successful transmission, the
station will unconditionally back off. This was ignored in [12-
14]. Furthermore, with the assumptions of traffic arriving in
Poisson process and the buffering of multiple frames [12, 13],
the transition probability from states having different retry
times to the buffer empty state should not be identical (which
unfortunately is assumed in [12, 13]) since the effective
arriving time interval of the next frame is different. In [12, 13],
the arriving time interval of the next frame is counted from the
arriving moment of the current frame to the successful
transmission moment of the current frame. It is different if the
current frame ends with different retry times. Consequently we
take the same assumption as in [11]: each device can buffer
only one frame and data arrives with a Poisson process. Thus
the effective traffic arriving interval of the next frame is
counted from the successful transmission moment of the
current frame and it is identical to the one-step state transition
period. Moreover related research on average service time in
IEEE 802.11 includes [9, 13-16]. All of them failed to take into
account the suspending events during the back-off process. In
this paper we extend the Bianchi's model by considering the
suspending events during the back-off process. Based on this
model we re-derive the average service time in IEEE 802.15.3.
The paper is outlined as follows. Section II briefly reviews
the IEEE 802.15.3 MAC mechanism. The extended analytical
model is presented in section III. Section IV validates the
model by comparing the analytical results with those obtained
through simulations. Finally, concluding remarks are given in
Section V.
II. IEEE 802.15.3 ACCESS MECHANISM
A. IEEE 802.15.3 MAC Protocol
In IEEE 802.15.3, the basic component is device and the
basic network element is piconet. One device is required to
assume the role of PNC (piconet coordinator). The time is
divided into superframes. The superframe is divided into three
parts as shown in Fig. 1: beacon frame, sent by PNC, used to
set time allocations and to communicate management
information for the piconet; contention access period (CAP),
accessing the channel in CSMA/CA manner, used to
communicate commands and/or asynchronous data if it is
present in the superframe; channel time allocation period
(CTAP), adopting the TDMA protocol, composed of channel
time allocations (CTAs), including management CTAs
(MCTAs). CTAs are used for commands, isochronous streams
and asynchronous data connections. MCTAs are a type of CTA
that is used for communication between the DEVs and the PNC.
Superframe #m 1 Superframe #m Superframe
the collision probability of a frame is constant and independent
of the number of retransmission.
We eliminate the assumption that at least one frame is
always available at each device. Furthermore we consider the
effect of frame errors introduced by channel noise. We also do
some simplifications in IEEE 802.15.3. PNC as a transmitter is
not considered since PNC has the absolute access priority
without the back-off process. The average frame delay of other
devices caused by PNC insertion is limited to PNC
transmission time. The device's waiting time (SIFS duration) at
the beginning of CAP is approximated with BIFS for simplicity.
A. The Modified Markov Chain Model
Consider a piconet with n devices and a PNC. Assume
each device can buffer one frame and there is a constant
probability q of at least one frame arriving per state, as
discussed in section I. Let bet) be the stochastic process
representing the back-off time counter and ret) be the
stochastic process representing the back-off stage for a given
device at slot time t. We introduce the state {(O,b(t))e} as
the back-off process after a successful transmission if the
buffer remains empty. The subsequent buffer empty state is
denoted as {(-1,0)} . u is the probability of channel sensed
busy, and P, is the frame-error probability. Different from
previous models, the suspending probability S during back-off
is supplemented. The modified markov chain model is depicted
in Fig. 2.
Next we analyze the back-off and transmission process in
detail. If the buffer has one frame to transmit, the device will
enter the back-off process after BIFS idle duration and the O-th
back-off timer is CWo' Note that only when the channel is
idle for a can the back-off timer decrease; if the channel is
occupied, the back-off timer will suspend and stay in the
current state. The probability is 1- sand s respectively,
where S is the suspending probability, namely the channel
sensed busy probability U . The state transition period differs:
if it is an idle slot, it lasts a ; if the slot is occupied, it can be a
successful transmission slot, a failed transmission slot due to
frame error or collision. Transmission is attempted when the
back-off timer reaches zero. The transmission is unsuccessful
with probability p due to collision or frame error; or it is
successful with probability 1- p . In the model p is the
transition probability from one row to the next row. In the
former case the next back-off and retransmission cycle begins
again until the transmission reaches the retry limit. In the latter
case the device either has a new frame to transmit with
probability q, or enters the empty back-off process {(O,k)e}
with probability 1- q if the buffer remains empty after a
successful transmission. As to the state {(O,k)e}, if no frame
arrives, it switches to state {(O,k-l)e} or remains in the
current state because of suspending; if a frame arrives, it
CTA n
Channel timeallocation period
Figure I. IEEE802.15.3 Piconetsuperframe
period

Deacon Conrention

III. ANANALYTICAL MODEL OF IEEE 802.15.3 CSMA/CA
In this section we present an analytical model of IEEE
802.15.3 based on Bianchi's model [7]. We impose some
assumptions similar to [7]: no hidden devices are considered;
B. IEEE 802.15.3 CSMA/CA Mechanism
The device first waits for back-off interframe space (BIFS)
duration, from when the medium is determined to be idle
before beginning the back-off algorithm. At the beginning of
the CAP, the device may begin the back-off algorithm a short
interframe space (SIFS) after the beacon transmission. Suppose
the device selects BO; as the back-off timer in the i-th back-
off process, then BO; =a x CW; , where CW; is the i-th
back-off window, CW; =rand[O, W; -1]
CW
min
S W; S CWmax and a is the slot time. For the first
transmission attempt of a frame, i is set to zero and W; is set
to CWmin' With the retry count i increments, W; is doubled
up to CWmax after the m'-th collisions,
CWmax =CWmin X 2m' Once W; reaches CWmax ' it will
remain at CWmax until it is reset. W; is reset to CWmin
either after every successful transmission or when retry count
reaches the retry limit m . The back-off timer BO; is
decremented only when the medium is idle for the entire
duration of slot time; it is suspended when the channel is busy.
The device resumes to continuously decrement the back-off
timer until the channel is measured idle for a BIFS. The device
may transmit the frame at the moment that the back-off timer
expires. After correctly receiving a frame in the destination
device, a positive ACK is sent to notify the source device after
SIFS. If the ACK is not received, the source device assumes
the transmitted frame is collided, and then it schedules a
retransmission and re-enters the back-off process.
948
s
,
I-s
I-s
I-s (m'+ I,W
m
, -I )
( I-s)( I-q/.-----L----..
- - --{ (O,Wo-l)e
s(l-q)
sq
I-s
I-s
s
t
s
(m',2)
(m'+1,2)
s
s
s
(m, l)
(m', I) .. I-s
I-s
I-s
I-s
( I-p)q
(m', 0)
(l -P)(l -q)
p
( I-p)q
(m'+ I,O)
(l-p)(l-q)
p
PI
q
(m.O)
Figure 2. A modified markov chain model for the back-off process in unsaturated traffic conditions
switches to state {(O,k)} or {(O,k-1)} similarly. The
transition probability is (1- q)s , (1- q)(1- s) , qs and
q(1- s) respectively. The transition probabilities for the
states {(O,k)e} and {(-1,O)} are identical: whether there is
a frame to come with the probability q , whether the channel is
sensed idle with the probability 1- u , whether the
transmission is successful with the probability 1- p . From
the above analysis, we can obtain the one-step transition
probabilities.
{
P{(i,k)I(i,k)}=s kE[I,W; -1] iE[O,m] (1)
P{(i,k -1) I(i,k)} = 1- s k E [I ,W; -1] i E [O,m]
1
P{(O,k)'I(O,k).}= s(1-q) kE[I,Wo-I]
P{(O,k-I).I(O,k).}=(1-s)(1-q) kE[I,Wo-I] (3)
P{(O,k) I(O,k)e} = sq k E [I ,W
o
-1]
P{(O,k -1) I(O,k).} = q(1- s) k E [I,W
o
-1]
P{(O,k) I(O,O)e} = qu kE[O,Wo-I]
W
o
P{(1,k)I(O,O)e}=q(1-u)p kE[O,Jf;-I] (4)
Jf;
P{(O,k)eI(O,O).} = p) k E [O,W
o
-1]
o
P{(-I,O)I(O,O)e} = I-q
P{ (O,k) I(i ,0)} = (1- p) q k e [O,W
o
-I] i e [O,m-I]
W
o
P{(O,k)el(i,O)}=(I-p)(I-q) ke[O,Wo-l] i e [O, m- l] (2)
W
o
P{ (O,k) I(m,O)}=.!L k e [O,W
o
-I]
W
o
P{(O,k)eI (m,O)} = I-q k e [O,W
o
- I]
W
o
P{(O,k) I(-I,O)} = qu kE[O,Wo-I]
W
o
kE[O,Jf;-I] (5)
1
P{(O,k)eI(-I,O)} = p) k E [O,W
o
-1]
o
P{(-I,O)I(-I,O)} = I-q
949
w= {CWrnin 2
i
i E [O,m']
1 "
CW
rnin2
m
iE[m+l,m]
(6)
E[ L
e
] be the average length of the longest frame payload
involved in a collision.
Pe is the probability that a frame encounters a collision.
The probability that a device transmits a frame in a slot
time can be evaluated as:
Note that the above average frame delay is derived
provided that this frame is not discarded, so the weighted
ffici b .c. T peN = i) d h f h
coe ticient erore di IS c an t e sum 0 t e
1- r'"
T
f
is the average failed transmission time.
C. Frame Discard Probability
The frame discard probability P
d
is the probability of
consecutive m+1 unsuccessful transmission attempts.
r, =pm+l (15)
D. Average Frame Delay
We denote as T
d
the average frame delay from the
moment the back-off process is initiated until the frame is
successfully transmitted. Let Ne be the number of collisions
experienced before a successful transmission, T
di
be the
average frame delay at the condition that N
e
is i . Then T
di
can be divided into four parts: i times back-off process, i
times failed transmission, i +I-th back-off process and the
successful transmission, then the average frame delay is:
{
I: = BIFS +PHYprelhdr +MAC
hdr
+E[L]+8 +SIFS +ACK+8
(14)
1;; = BIFS +PHYprelhdr +MAC
hdr
+E[L
c]+8
I: = BIFS +PHYprelhdr +MAC
hdr
+E[L]+8
m
LP(N
c
=i)
coefficients is ;=0 , equaling to 1.
1- r'"
T -i.t +T](16)
d LJ 1- m+l di LJ 1- m+l LJ b ]; f s
i=O P i=O P k=O
(9)
(8)
(11) P; = 1- (I - r )n
P; =u=s=I-(I-r)n-1
m
t =Lb;,o +Q(I-u)[b(o,o)e +b_l,o]
;=0
where P{ (iI' k
l
) I(i
o
, k
o
) } is the short notation of
P{r(t +1) =il,b(t +1) =k
l
Iret) =io,b(t) =k
o}
Combining equation (1)-(10), a nonlinear system with one
unknown T is formed. Due to the space limitation, we refer
the readers to [7] for the detailed derivation process. We
consider not only the unsaturated traffic condition but also the
device suspending event, so the process to solve the equations
is much more complicated than previous models. We can fmd
the solution by using the symbolic math toolbox in MATLAB.
B. Throughput
is the probability that at least one transmission occurs.
is the probability that exactly one device transmits,
conditioned on the fact that at least one device transmits
Let bi,k ' bU,k)e and b_l,o represent the stationary
distribution and normalize them to 1, we can solve bo,o.
nr(l- r)n-l
p=----
s
nr(l- r)n-l
I-(I-r)n
(12)
(17)
(18)
The normalized throughput in an error-prone channel is
= PsP
tr
(1- P
e
) E[T
L
] (13)
Snorm (1- Ptr)O" +PtrP
s
(1- P, )T
s
+P; (1- P,)T
c
+PtrPsp:r.
If P, is zero, equation (13) agrees with [7] and [11].
is the average frame payload transmission time; T, '
T; and T
e
are the average times that the medium is sensed
busy due to a successful transmission, due to a collision and
due to transmission errors respectively. Let 8 be the
propagation delay, E[ L] be the average frame length, and
T
bk
is the back-off time during the k+l-th back-off process.
Let Tstate be the one-step state transition period in Fig. 2,
Tdeere be the average time for the back-off timer to decrease by
1, namely the arriving time from state (k, j) to (k, j - 1) ,
N
decre
be the number of steps staying in state (k,j) before
jumping to (k, j - 1)
w: -I
r; =T Tdecre
950
Pis is the probability of exactly one transmission from the
n -1 remaining devices :
I
.
I
I
I
,
I
30 35 40
I
l
H.Wu model in [7]
12 c _ K.Duffy model in (11) f-----+--+---"-I,----
- our model
.v simulat ion

l I
20 I
18 I
;j
i 16
e
I
unsaturated traffic model is based on the model in [II],
whereas [II] did not take into account the effect of device
suspending and did not address the average frame delay. [16]
gave the mean and standard deviation of the frame delay in
saturated traffic conditions using one-dimensional markov
chain without considering the suspending events . We compare
the throughput analysis with [7] and [II] in Fig. 3 and compare
the average frame delay with [15] and [16] in Fig. 4 to
demonstrate the effect of considering unsaturated traffic
conditions and suspending events in performance modeling.
All of the above works assume the channel is ideal, so the
frame error rate P
e
in our model is set to zero for fair
comparison purpose.
In Fig. 3, the model in [7] assumes a saturated traffic
condition and q is set to 0.1 in [II] and our model. All curves
demonstrate the same trend: with the increased number of
devices the throughput increases and achieves a maximum and
then decreases . Once all of the bandwidth is occupied, the
throughput can not increase with the number of devices . On the
contrary the throughput will decrease with the increase of the
collision probability. Comparing the models in [7] and [II], we
can observe that the maximum throughput is reached with a
smaller number of devices in the saturated traffic case than in
the unsaturated traffic case. The larger the q is, the earlier the
throughput becomes saturated.
As shown in Fig. 3, our analytical model matches the NS-2
simulation results much better than the others. When
suspending event is not considered, the device at the back off
state will jump faster to the state to transmit a frame than
otherwise. This means the probability to transmit a frame in a
given slot is larger when not considering the suspending event .
If the network bandwidth is not occupied fully, larger
probability to transmit a frame results in larger throughput. As
the network becomes saturated with the increased number of
devices, a larger probability to transmit a frame will lower the
throughput. This suggests the performance analysis in [7] and
[II] which ignores the suspending event during the back-off
process is underestimated for large number of devices.
15 20 25
Number of de";ces
(19)
(21)
IEEE 802.15.3 PHY AND MAC PARAMETERS TABLE!.
Symbol Value Symbol Value
a (J.1S) 17.273
CW
min
8
SIFS (J.1S)
10
CW
max
64
BIFS (J.1S)
17.273
ACK(bytes)
10
PHY
n rJJ.1S
)
17.455
b(J.1S)
1
ntv.; (J.1S)
0.727
L(bytes)
256
MAC
hdr
(J.1S)
3.636 m 3
(n -1)T(1- Ty-2
Pis =
Pc
After some algebra operations, (16) becomes
Tdecre=i:: = r.: +'F.tatel
1=0
= fs/(1-S)'(I+I)T = 'F.tate
1=0 state 1- S
During the one-step state transition, the channel may be idle
or occupied because of successful transmission, frame error or
collision by the remaining n -1 devices
Tstate =(1-U)a-+UPls(1-P
e
)T, +uP,sPeTe +u(1- p's)T
c
(20)
IV. NUMERICAL RESULTS
According to [1], the specific network configuration used in
our simulations is listed in table I. The base rate is 22Mbps and
the data rate is 55Mbps. In order to verify the accuracy of the
proposed model, the results obtained using the IEEE 802.15.3
network in NS-2 are also plotted in the following figures .
Related IEEE 802.15.3 simulation environment in NS-2 is set
up by Mustafa [17]. The superframe size is set to 15
microseconds and it consists of only the beacon frame and
CAP to avoid the implication that CTAP may pose on the delay
measurement.
T 2
m'c w 2( m m') m+2
= slate .{ minP P - P - P +P
2(l-s)(l- p
m
+
1
) 1- p
-CW . p[I-(2p)m'+'] +2m'CW . (m-m' +2) (22)
nun 1- 2p nun
1 (m+l)pm
-CW . -m-l}+T .p[-- ]+T
nun f 1 1 m+l S
-p -p
Figure 3. Throughput for different number ofdevices
The models in [7] and [15] are both based on Bianchi's
model but considering the retry limit in saturated traffic
conditions. [7] derived the expression for the throughput and
[15] derived the expression for the average frame delay. Our
Fig. 4 shows the average frame delay for different models.
Again, our analytical model matches the NS-2 simulation
results much better than the others for both saturated and
951
4.5 - - - - - - - - -- - -- - -- - -- - -;
Figure 4. Average frame delay for different number of devices
REFERENCES
[I] Part 15.3: Wireless medium access control (MAC) and physical layer
(PHY) specifications for high rate wireless personal area networks
(WPAN), IEEE Std 802.15.3, Sep. 2003
V. CONCLUSIONS
In this paper we propose an extended analytical model to
study the performance of the IEEE 802.15.3 CSMAlCA
mechanism. Both the back-off process after a successful
transmission in unsaturated conditions and the suspending
event during the back-off process are considered. The
expression for the average frame delay is re-derived based on
the modified model. The explicit analytical model has been
validated to be in agreement with the computer simulations.
[2] X. Shen, W. Zhuang, H. Jiang, and 1. Cai, "Medium Access Control in
Ultra-Wideband Wireless Networks", IEEE Tran. on Vehicular
Technology, vo1.54, no.5, pp, 1663-1677, Sep. 2005
[3] Part 11: Wireless LAN Medium Access Control (MAC) and Physical
Layer (PHY) Specification: Higher-Speed Physical Layer Extension in
the 2.4GHz Band, IEEE Std 802.11b, 1999
[4] F.Cali, M.Conti, E.Gregori, "Dynamic Tuning of the IEEE 802.11
Protocol to Achieve a Theoretical Throughput Limit", IEEE/ACM Tran.
On Networking, vo1.8, no.6, Dec. 2000
[5] G.Bianchi, "IEEE 802.11 Saturation Throughput Analysis", IEEE
Comm. Letters, vo1.2, no.12, Dec. 1998
[6] G.Bianchi, "Performance Analysis of the IEEE 802.11 Distributed
Coordination Function", IEEE Journal on Selected Areas in
Communications, vol.l8, no.3, pp.535-547, Mar. 2000
[7] H.Wu, Y.Peng, K.Long, S.Cheng, 1.Ma, "Performance of Reliable
Transport Protocol over IEEE 802.11 Wireless Lan: Analysis and
Enhancement", in Proc. IEEE Infocom, New York, USA, pp.599-607,
2002
[8] y'Xiao, "Performance analysis of IEEE 802.11e EDCF under satuarion
condition", in Proc. IEEE ICC, Paris, France, pp.170-174, June 2004
[9] J. Tantra, C. H. Foh, and A. Mnaouer, "Throughput and delay analysis
of the IEEE 802.11e EDCA saturation", in Proc. IEEE ICC, Seoul,
Korea, pp. 3450-3454, 2005
[10] L. Xiong and G. Mao, "Saturated throughput analysis of IEEE 802.11e
EDCA", Computer Networks, vo1.51 , no.11, pp.3047-3068, Aug. 2007
[11] K. Duffy, D. Malone and DJ. Leith, "Modeling the 802.11 distributed
coordination function in non-saturated conditions", IEEE Comm. Letters,
vo1.9, no.8, pp.715-717, Aug. 2005
[12] W.Lee, C.Wang, and K.Sohraby, "On use of traditional M1G/I model for
IEEE 802.11 DCF in unsaturated traffic conditions", in Proc. IEEE
WCNC, Las Vegas, USA, pp.1933-1937, Apr. 2006
[13] O.Rodolfo, B.Luis, P.Paulo, "Modelling Delay on IEEE 802.11 MAC
Protocol for Unicast and Broadcast Nonsaturated Traffic", in Proc. IEEE
WCNC, HongKong, China, pp.463-467, Mar. 2007
[14] LJing, N.Zhisheng, "Delay Analysis of IEEE 802.11e EDCA Under
Unsaturated Conditions", in Proc. IEEE WCNC, HongKong, China,
pp.430-434, Mar. 2007
[15] P. Chatzimisios, A. C. Boucouvalas, and V. Vitsas, "IEEE 802.11 frame
delay - a finite retry limit analysis," in Proc. IEEE Globecom,
SanFrancisco, USA, pp.950-954, Dec. 2003
[16] T.Sakurai, H.L.Vu, "MAC Access Delay of IEEE 802.11 DCF", IEEE
Tran. on Wireless Communications, vol.6, no.5, pp.1702-1710, May
2007
[17] R.Mangharam, M.Demirhan, R.Rajkumar and D.Raychaudhuri, "Size
Matters: Size-based Scheduling for MPEG-4 over Wireless Channels",
SPIE Conference on Multimedia Computing and Networking, 2004
40 35 30 15 20 25
Number of de"";ces
10
P.Chatzi mi sios model in [15]
I
-
T.Sakurai model in [16J
;a-: J
" ----a- our model , q=0.99 ,rr .. I
--
/." " .
--:-- our rnodel.q-Oil ,2'" : .
I 2
,..,;--
"
sim, saturated (q=l ) . ..
sim, unsaturated (q=0.1)
r" ./-'
,;rr"'.
..,. .
/ ';: : -
-

..,.,, '
--
I
/,":


I .'
" if / "
I / . '
-"/ , ,
I

r r r r r r
I
0.5
3.5
4
.. 3
,
'"
2.5
Q)
E
2
Q)

1.5

unsaturated cases. In particular, let us compare the models in

[IS] , [16] and ours in saturated conditions. As there exist
suspending events, the back-off timer has more chance to stay
in the current state. The larger the number of devices, the
higher probability to stay in the current state. The average
frame delay does not become saturated as suggested by the
other two models with the increased number of devices even
for a finite retry limit. This illuminates that there can not be too
many devices in a piconet in order to maintain a low frame
delay.
952